Contributions to rotation invariant character recognition

Optical character recognition of machine-and hand-written numeral images has been improved by new methods for the rotation and translation invariant recognition of images using the rotation-invariant properties of an othogonal two-dimensional transform, the Zernike transform. Two algorithms for the estimation of the rotation invariant coefficient magnitude have been developed and tested. In addition, a new rotation normalisation procedure is presented, which performs a rotation of the image into a standard position. In the case of machine-written digits, the magnitudes of the Zernike coefficients are used as features, and classification is performed with the Condensed Nearest Neighbour Network. For hand-written digits the rotation normalisation procedure is applied, and images are classified with a Multilayer Perceptron resulting in a recognition rate of 96 %. (WEN)Available from TIB Hannover: RR 4652(8) / FIZ - Fachinformationszzentrum Karlsruhe / TIB - Technische InformationsbibliothekSIGLEDEGerman